Rosy

[cloud rivet]

just for fun

oh wow, are you looking to do video games for a job?

[gaunt drum]

yeah, as an indie game dev

[true moon]

what kind of games are you making

[cloud rivet]

good luck, that's awesome

[gaunt drum]

nothing unfortunately, I got many ||burnout sessions|| and NIHd a lot since 2024. had unsuccessful team experience. at this point I'm just too deep and prideful to give up and just trying to scramble something

due to where I live my expenses are not that high, so I am able to take the risk. I wouldn't risk it if I were in europe mainland or hell, us

[gaunt drum]

but currently since I abandoned my last NIH project a week ago, a horror game using godot 😂 but I suck at game design and been trying to figure out how I can make something actually fun

[true moon]

horror is a tough one too lol

the atmosphere has to be just right

what genre of game do you enjoy playing the most

[gaunt drum]

unfortunately coop experiences

[true moon]

coop like online multiplayer stuff, or story games that are coop

[echo crystal]

i am making a horror game too, still building the basic engine tho

[gaunt drum]

2-4 player games, like those coop job simulation games, or coop horror games like lethal company, phasmophobia, survival games like project zomboid, valheim. there isn't a specific genre that I can pinpoint into

[true moon]

ah yeah

zomboid is great although I wouldn't recommend making your own by yourself

[cloud rivet]

you got this

[gaunt drum]

oh, hey 👋 that I mentioned failed team initiative was also a horror game, mostly a phasmophobia clone, or a phasmophobia variant we wanted. then I got bullied myself into not using an engine, did some rust NIH stuff with wgpu. then zig & odin vulkan stuff recently. there are some breakaway tries to using engines but NIH was strong. I mostly lacked a clear project goal though, so I randomly chose technical issues to pursue instead

like "this is going to be a 3d game that I don't know yet", that vague

[cloud rivet]

I love psychological, dark and ominous horror movies, I haven't ever played a horror game although Metro Exodus and HL2 kind of have horror elements, I played Alien Isolation for a bit, I guess that's kind of a horror game, but probably none of those are what people think of when they think horror game

[echo crystal]

real

[gaunt drum]

but recently, after the talk with bjorn a week ago, around the same time I saw something on twitter about a local indie/publisher I decided to go all in on engines. yeah I still think it sucks and I hate it but I don't have many choice atm. it is what it is, time to accept it

[cloud rivet]

that's what I would do

[echo crystal]

psychological horror is cool

[cloud rivet]

if I ever want to make game in the future, I would use Unreal Engine and I would find 2 talented and committed artists to partner with who also wanted to make a small game. If I can't find that I'm not making a game.

[gaunt drum]

alien isolation was so scary, I watched a friend playing it

[cloud rivet]

it's really fun, but I'm not good at video games

and so I cannot do well

I am really greatful for bindless in vk, because I'm currently writing my descriptor set initializer and I had forgotten how horrible the descriptor set API is

[echo crystal]

same

[cloud rivet]

push constants working

now I have to write a bunch of memory allocators and math code before the pixels can change in an interesting way

[cloud rivet]

my videos always demo a full rebuild to launch as cl doesn't do incremental builds and it's a unity build anyway :P

not that there's a lot going on anyway

it's probably around 1.5K loc of mostly create info struct spam

there's probably a raddbg cli arg to just start the program that would make it faster, since a lot of the initial delay is just me staring at the screen waiting to press F5 on the keyboard

[cloud rivet]

did some research today for how to blit to a win32 window handle, I think it will take me a few tries before I understand how to even draw anything

also I guess I will start on a bit of math first before I start on an allocator, because I can push an orthographic projection via a push constant, and know that my math is working

then resizing the window won't skew the triangle

and that math will help me as a reference the math for the software rasterizer when I get to that

so I have two GP projects now, my C vulkan renderer and my win32 software rasterizer

only the software rasterizer is just restarted every day I do it

and a third is the computer enhance course sorta

[bronze socket]

have you ever tried that minimal windowing lib for writing diy software rasterizers

blanking on the name even though I've used it in the past

[brisk chasm]

nanokatze has a gist pinned in their gists about resizing vulkan stuff

[cloud rivet]

I haven't ever done a software rasterizer

I have an idea for how it should work based on my understanding of the math and the graphics pipeline

my window resizing works fine, it's just there's no projection of any kind so the NDC coordinates are just naively sent as screen coordinates and that's why I got the tall boi triangle up there I shared recently

I just need to write some math code to do projection and send it via push constant until I have an allocator and can actually create buffers on the GPU

I'm going to fill in all the host allocator stuff in the vulkan api too though, not just for the gpu allocation

those optional allocator parameters when creating things during init etc

[astral hinge]

I always forget the point of those allocator callbacks in Vulkan

I think it's for debugging drivers

[cloud rivet]

I just want to do it for fun

and to be responsible for all of the memory my process allocates

as the spec says, none of that allocation is happening in the critical paths

[astral hinge]

It would be interesting to see when the allocator is invoked in the driver

[cloud rivet]

hrm true, I'm curious as to when it wants to realloc

I don't know how hard the memory work that VMA usually does will be

Rosy used Volk, VMA, Dear Imgui, fastgltf, the fbx sdk, and SDL3, and a json library for the level state, and the C++ std lib, which offers a lot

I don't have any of those things anymore

oh and I used the nvidia image library that did all the compression

and mikktspace

hrmm

I got a lot of work to do

[brisk chasm]

so you are writing your own vulkan loader too?

[bronze socket]

the work for VMA is only as hard as you need your allocator to be

I prefer not having to think about it personally, it's just not that big of a deal to learn about tbh

[brisk chasm]

yeah

same

[cloud rivet]

yes I am writing everything from scratch

I can't write the OS or C runtime or Vulkan stuff from scratch, but everything else

we'll see how it goes with images,

I'd like to load the khronos sample gltfs and render them and and their images correctly, but that's not for a long time

[wraith urchin]

Skill issue.

[astral hinge]

https://tenor.com/view/bait-thats-bait-tom-hardy-mad-max-gif-5055384

[echo crystal]

making ur own OS

[true moon]

The GPU driver stack generally makes that infeasible unfortunately if you want to use the GPU

Maybe some insane mf has written a bare metal driver for AMD or something

[echo crystal]

u could make a toy gpu

https://www.furygpu.com/

pretty cool

[astral hinge]

you order it and you get a GPU covered in fur

[gaunt drum]

pull a jonathan blow

https://tenor.com/view/doja-cat-star-wars-gif-25078126

[cloud rivet]

yeah I'm not a jonathan blow, I didn't know what SRGB was like a year ago

[echo crystal]

just keep doing what u are doing now

learning stuff + having fun =

[cloud rivet]

I nearly got a blit to my window working in today’s video. After the recording I wanted to see how close I was and got my bytes to color the window within a few minutes. Huge progress from yesterday

[brisk chasm]

fopen /dev/dri/card0

ah i forgot you prefer windows for some reason

[cloud rivet]

What is the fopen referring to?

[cloud rivet]

Is that how you blit to the window? Open the window buffer as a file?

[cloud rivet]

ah looks like a video card thing

[brisk chasm]

dri is linux' way to talk to the gpu

and on lunix everything is a file

https://www.youtube.com/watch?v=B8xHWwCuxKA&list=PL980gcR1LE3IwpeihtCRVFK6maXqBb6AN

[bronze tendon]

I had to play with that at work once. Kinda funky how you can just sorta jack whatever's on screen.

[broken fog]

to the gpu driver, right?

[cloud rivet]

I was thinking the trying to build a from scratch to rasterizer in one hour thing was going to be my side project to my main project the C vulkan renderer, but instead it has become the main project, and I think I will only have time to work on my vulkan renderer on the weekened, which is fine. I'm having a lot of fun with the daily task of making more progress. It's challenging too, and feels like real learning.

[true moon]

There's a reason most good graphics courses are built around software raster

At least ones focused on real time foundations

[cloud rivet]

I haven't even gotten to the raster yet lol, I'm still having trouble with basic C and the win32 api, but I am thinking through that problem

[bronze socket]

one of my favorite projects of learning vulkan was a software rasterizer that evaluated in quads like a GPU and interpreted spirv

[cloud rivet]

I really have to condense everything down and get faster on things as I make progress, I'm going to add time stamps to my video for each milestone to measure how long it took me to get through each step

[bronze socket]

really learned a lot about spirv there

[cloud rivet]

like just opening up the window took up the full first video

and on the second one it took 30 minutes

and then that'll just keep shrinking

it's also pretty cringe to watch the videos myself, I see all the mistakes I made after I made them, seems obvious afterwards

but with previous projects there any new thing I did was always hard as I did it, but then eventually I moved on to new things I actually would forget how I solved those problems

having to do it again day after day is a really great way to imprint things onto my brain

it's fun though

I don't care if anyone ever watches the videos it's kind of exactly what I want to do right now

but a part of me thinks if I get good and fast enough that the dramatic difference between the first video and the video when I'm actually making good progress will be interesting to people

went from struggling to open a window to building something that does something cool

[cloud rivet]

I am not sure if I explained the thing

Every day I start with a completely empty project and the goal is to get it done in an hour

[dark saffron]

@cloud rivet I somehow doubt your crashes are due to validation fails

usually the VVLs complaints are irrelevant to the drivers

[cloud rivet]

I see, hrm. I will see if I can investigate the crash more. We can assume it’s not vcc related since I don’t have any direct evidence.

I bet there’s a bug in my arena allocation

It’s an access violation and the size of the vcc spv is much larger than what slang generates

If that’s not it I will learn the spirv spec and try to find a minimal way to reproduce

Correlation is not causation moment

I have learned from my C shenanigans that C++ shielded me from how horribly you can break not just your application but also your computer with bad bugs. Just absolutely bizarre behavior after the application exits.

I think yesterday I put Windows into a bad state with UB

It’s pretty cool

I did that during recording. So there’s video evidence

I figured out after what code of mine was causing all that, not how, but it was obviously UB and fixing it fixed the mess it was making

[wraith urchin]

You can do this with C++ too if you try

[cloud rivet]

I love computers

[dark saffron]

the size of the generated code is not a useful metric

also it should have gone way down with the latest commits

it might very well be a Vcc issue, or an nvidia bug, or some weird edge case where both meet

the fact you're using shader objects might have something to do with it

i haven't tried that codepath at all

[cloud rivet]

I will focus on reproducing this weekend

[dark saffron]

thanks!

[cloud rivet]

no problem, will be fun

I could eliminate SO by using the alternative and see if that works as a first step

I will try that

[cloud rivet]

one of my videos got a like from someone and the views 10'xed, then the next did not and it went back to next to none.

effectively no views, which is fine

just interesting, I think will try to make the videos better though, they are cringe

like why make a video at all if they're going to be bad

I'm going to build like a premiere template or something to make video editing fast, I can't spend a lot of time on that as I hate it, not editing the videos at all right now

I can spend a little bit of effort though and it will probably produce a nice outcome

I really like making the videos and doing this exercise, but it does kind of make my skills and problem solving performance, or lack of, very public

nobody's watching them though, so it's fine

it's cool to see how I made progress from one video to the next

which is why I am doing that

[wraith urchin]

Wait did you link the videos here?

[cloud rivet]

no I haven't shared them, they're very bad right now

they're just going into the YT algo without me marketing anything

[wraith urchin]

You should link them here lol

We want to see them

[cloud rivet]

well, I don't want you to, but if you want to, I link my YT in my profile, but they are super cringe atm, and will get better

maybe you have some feedback about what I could be doing better. One problem is that I live in a very small place, I do not have my own office, so my family can sometimes be heard in the background

also the first video has no audio

because I didn't know what I was doing

the audio on the second video is also very bad, I bought a new headset

[wraith urchin]

Noise gate

[cloud rivet]

what is a noise gate

I will search

[wraith urchin]

are you using obs?

[cloud rivet]

yes

[gaunt drum]

it blocks voice below certain freq

[wraith urchin]

Not exactly

[gaunt drum]

👀

[wraith urchin]

Obs has a noise gate in the audio filters

[cloud rivet]

oh nice

[wraith urchin]

It lowers the volume when you are not speaking

[cloud rivet]

I saw premiere had audio options too, so I was going to look into those too

like on the weekend figure out a. nice template

[wraith urchin]

It's like the opposite of a compressor

[cloud rivet]

ah

[wraith urchin]

Another tip. Get as close as possible to your mic

[gaunt drum]

oh this one

[cloud rivet]

it's a headset mic

it's very bad

[gaunt drum]

I was using nvidia's noise reduction thingy

similar to discord's one in function I think

[wraith urchin]

I'll check out your vids in a little bit and see if I have more suggestions

[cloud rivet]

thank you

I was thinking a mic would pick up even more of the surrounding sounds than a headset mic

is that incorrect?

the headset already picks up a lot of ambient noise

maybe there's an attachment or something for that

I don't really want to invest too much into this until I know I want to keep doing it too

[bronze socket]

this kinda sounds like speedrun training more than anything too useful for programming tbh

[cloud rivet]

yeah kinda, but I learn from it

but it is definitely geared towards taking shortcuts

I'm not trying to be an educator, nobody should learn from me lol, I don't know what I am doing

you can't get good results without being effective, even if it is speedrunning

[gaunt drum]

as long as you are having fun, that's the only thing that matters unless it is your job

[cloud rivet]

yeah

it is fun

it has some moments of stress when things go unexpected

I'm just going to keep doing it until I want to do something else idk

[wraith urchin]

Not necessarily. Mics have different polar patterns, the one you want is cardioid. With headsets its usually a tiny electret mic, which are omnidirectional unless the headset designer carefully designed the housing to make it cardioid which is hard to do in a tiny enclosure

Having a cardiod mic close to your face also gives a natural bass boost which gives people the "radio voice" which might be desired

@cloud rivet background noise isn't too bad, but your headset mic is off axis, so it sounds very tinny

the mic should be basically right in front of your mouth

it could also just be a mediocre headset

Are you willing to buy a new mic? You don't have to spend all that much to get a big improvement

[cloud rivet]

yes

I would love to have a better mic

[cloud rivet]

if you have a recommendation for a good mic I would appreciate it! 🙏

[wraith urchin]

how much do you wanna spend?

[cloud rivet]

< $300 ?

idk

[wraith urchin]

yeah you don't even need that much

[cloud rivet]

I don't spend money ever, my biggest expense is coffee

[wraith urchin]

the AT2020 is a good workhorse and a solid mic

[cloud rivet]

ok, I will look a that one

[wraith urchin]

It comes in two versions a USB version and an XLR version

[cloud rivet]

I don't know what XLR is, does that require hardware?

[wraith urchin]

if you get the XLR one you also need an audio interface

[cloud rivet]

hrm

is it valuable to get an XLR version?

[wraith urchin]

It gives you more control, and more options if you wanna upgrade later

[cloud rivet]

thanks for taking a look!

your audio sounds great in your stream

[bronze socket]

ah the good ol' red box

[wraith urchin]

I use a focusrite 18i20, which is like a bigger version of that. and an AT3035 which is like an older and slightly higher end version of the AT2020

so if you get the focusrite and the AT2020, it will be pretty close to my setup

[bronze socket]

idk about mic stuff but you'll definitely want that audio interface if you ever decide to get into electric guitar

[wraith urchin]

https://www.amazon.com/Adjustable-Microphone-Suspension-Broadcasting-Voice-Over/dp/B00DY1F2CS?crid=1SNNP1D1VMINH&dib=eyJ2IjoiMSJ9.BbO6sNKWP6JELG25w-Zwdlc3QqIMQLZH-U-5mk2yVKB3_t7OvpYqF_DnPN-ZmbSrTBAIDJtqNEqdIeWcyBn8_qRYQUPpojHJSbgSjKcQqH8drb4ZczAL3Kjn2nDC1IYpTjpakkavwY9c7_IVMjKBqXhQsT8qKlcSkTKVQzKsq1vp8fKH-nLK9vc91FOvbNYYTXj49xLBoVvSBVS2S6YAe7SrVMbzKjjAgR2eff0821wxAxo2D4RuN3zNhTw_VIDz0gsqcQ-oEfgzruuR1ktWlaaC9e3a2i_ek19HRTWHdwg.pjZXXzmcdwvo7ez0ACkPvJ1Dgh63BkET2cw0lvtjK3Y&dib_tag=se&keywords=neewer%2Bmic%2Bstand&qid=1752772566&sprefix=neewer%2Bmic%2Bstand%2Caps%2C122&sr=8-5&th=1

this is the boom arm I use

[wraith urchin]

most electric guitar you hear on recordings is actually just a mic on the guitar cabinet

usually a shure sm57

[bronze socket]

yeah for studio stuff, for hobbyists who just plug their guitar into their PC they use an audio interface

[wraith urchin]

Sigh, future generations will never know the hearing damage I suffered doing recordings in the Toronto punk scene

[cloud rivet]

thanks, bought it

[wraith urchin]

focusrite and at2020?

solid combo

did you get the boom arm?

[cloud rivet]

yes

[wraith urchin]

sweet

that should be a very reliable setup

[cloud rivet]

thank you!

I am not going to play any instruments

[wraith urchin]

no worries, feel free to hit me up if you run into any issues

Oh yeah one more bonus for using the interface instead of the usb version is you get zero latency monitoring

so you can hear your voice coming in through the mic before it goes into the computer

[cloud rivet]

nice

[cloud rivet]

finally making some progress beyond the win32 C stuff

[cloud rivet]

https://youtu.be/gzwosVfE-Zs

[wraith urchin]

Your thumbnail game is very strong btw

this face makes me wanna click

[cloud rivet]

LegendsOfTotalwar is a YT creator who makes videos solving other people’s failing Total War Warhammer games.

Anyway, his thumbnails are amazing. You have to know the game really well to get them, but they are often hilarious.

Being able to stick a joke into a thumbnail that your target audience gets is probably too tier thumbnail game. Would be cool to see if I can figure that out for my thing

I am not going to daily spam this channel with my YT videos btw. Maybe sometimes when I achieve a new milestone

Going to work on trying to get to the bottom of my vcc issues this weekend

[hushed creek]

bjorn be looking up to his name like a viking metalhead XD

cool

[cloud rivet]

:P

it's even worse when I have a beard

people can probably guess my name is Bjorn without me telling them

ok that's enough PFPs from me lol

[alpine basin]

I see why you're called Bjorn since you look like that

[wraith urchin]

Nominative determinism?

[cloud rivet]

goal for this weekend is to create pipeline shaders to see if that gets my vcc shaders to run, build out some math, build a premiere template of somekind and write some neovim code to make building bat files easier and faster

[cloud rivet]

the new mic, scarlett and stand work very nicely thank you!

[wraith urchin]

Awesome, can't wait to hear them in action

[cloud rivet]

anyway

awesome, I will shout out ASO :P

for all two viewers, one of which is you

[wraith urchin]

It sounds so much more professional now

[gaunt drum]

yeah it is good

unrelated, I envy you being ~15 years older than me and still having that perfect hairline

I too have long hair but I almost lost half of it

[cloud rivet]

I turn 50 in 17 months, maybe it will all fall out then

I don't even know how to create shaders without SO, I have to like go look that up

I think I sort of know

the only reason I switched to Vk was because some actual game developer convinced me that SO meant VK was now just "fancy opengl"

turns out AMD is just emulating SO and there's a new shader objects v2 draft that's not public yet

I asked on the AMD discord

Current situation is: RADV (Linux) has a "native" implementation of EXT_shader_object. AMDVLK (Windows) just emulates it on top of EXT_graphics_pipeline_library for now.

There is a more fleshed-out shader_object v2 spec in the works, but that won't be ready to talk about for a while yet.

I don't really have a lot going on in palinode yet so it's fine

I think maybe though I don't know if I need a GPU to do graphics and have fun tbh

maybe GPUs were a mistake

you know what's interesting is

[vagrant musk]

yeah I think I'm gonna let vulkan core spec mature a bit more before I go back to using it

[cloud rivet]

maybe I can use vcc code in my software rasterizer and in a shader

isn't Vulkan as old now as gl was when it hit its last version?

I don't know

and released it on June 30, 1992.

4.6[1] 31 July 2017

not even close

[vagrant musk]

vulkan: 10 years
opengl first release: 1992
4.6: 2017
eh in 15 years it will be

[cloud rivet]

yeah

so 2/3 of the way into opengl's lifespan

I think by the time vulkan is 15 AI will just on the fly generate all the frames

so no more graphics APIs :/

sorry

[vagrant musk]

the vulkan core spec is slowly converging to be closer to opengl's level of dynamicness
which is nice for me, considering I am sorta heavily used to dynamicness

[cloud rivet]

you might like Metal

[vagrant musk]

me with my exclusively windows computers

[cloud rivet]

same

[true moon]

Me with my diverse cast of computers none of which are macs

[alpine basin]

Me with two computers that run Linux

[cloud rivet]

I can't own more than one PC and one laptop. I could maybe, maybe have like a little mac mini or tiny little Nuc or something with linux on it

but I don't like having a server farm under my desk or in a closet or whatever like all my coworkers

man I get so nervous before I start recording and I have no idea why, it's just a recording

[true moon]

Are you doing any editing at all

[cloud rivet]

it's a timed thing so no I don't want any cuts

I think?

I'm not sure

I might

I have premiere

I figured out how to cut with it

I am going to play with it this weekend see if I can make a nice template

[true moon]

I was just thinking if you don't mind doing a tiny bit of editing I'd just start rolling but plan on not starting the video right away, give yourself a minute to get situated with the recording on and then begin

just to beat the nerves of the start

[cloud rivet]

that's a good idea

[true moon]

I guess on the other hand just starting immediately into the video does sort of force you to just accept what you get rather than giving yourself room to agonize over the perfect start

ripping the bandaid off

[cloud rivet]

yeah, I just try to leg to and deal with it, I kind of just want to not hide anything about how hard this is for me, or any struggle I have

also sometimes when there's like something not working right I get some anxiety about being able to figure it out, but I force myself to focus

[cloud rivet]

Shoutout to ASO ended up at the end, I forgot until the end, I'll remember to do it first next time. I will add it to my notes. Ended up having the mic too close to my face or the input too high or something and the video has unfortunate and annoying popping, I will get it right next time. :(

[silver slate]

High quality testing.

Yeah, when you don't have a pop filter you usually have to make sure you're talking somewhat past the microphone, instead of directly at it.

[cloud rivet]

vcc shader triangle 🎉

I also now have both shader objects and a graphics pipeline

#define push_constant          __attribute__((annotate("shady::io::392")))

going to try push constants in a minute and then if that works I can default to vcc instead of slang

and then I can share my math.c code both application and in shader

maybe I just do rasterization in compute and not use a graphics pipeline, tbh

it's kind of fun

then both projects are very similar and I can learn from both

[cloud rivet]

vcc push constant

my shader:

#include <shady.h>
#include <stdint.h>

float floorf(float) __asm__("shady::prim_op::floor");
float fmodf(float, float) __asm__("shady::prim_op::mod");
float sinf(float) __asm__("shady::prim_op::sin");

location(0) output native_vec4 out_color;

typedef struct {
    float t;
} pc_t;

push_constant pc_t pc;

// A naive triangle
vertex_shader void main()
{
    float pi = 3.1415926535897932385f;
    native_vec4 pos;
    int i = 0;
    switch (gl_VertexIndex) {
        case 0:
            {
                pos = (native_vec4){-0.5f, 0.5f, 0.f, 1.f};
                out_color = (native_vec4){1.f, 0.f, 0.f, 1.f};
                break;
            }
        case 1:
            {
                pos = (native_vec4){0.f, -0.5f, 0.f, 1.f};
                out_color = (native_vec4){0.f, 1.f, 0.f, 1.f};
                break;
            }
        case 2:
            {
                pos = (native_vec4){0.5f, 0.5f, 0.f, 1.f};
                out_color = (native_vec4){0.f, 0.f, 1.f, 1.f};
                break;
            }
        default:
            {
                // Garbage data detectable with Renderdoc
                pos = (native_vec4){7.77f, 7.77f, 7.77f, 7.77f};
                out_color = (native_vec4){1.f, 0.f, 1.f, 1.f};
                break;
            }
    }
    float x_move_interval_ms = 10000.0f;
    float ms = floorf(pc.t * 1000.f);
    float test_1 = fmodf(ms, x_move_interval_ms); // value from 0 to 1000;
    float test_2 = (test_1 / x_move_interval_ms);
    float test_3 = test_2 * 2.f * (float)pi;
    float s_x_pos_mod = sinf(test_3);
    pos.x *= (s_x_pos_mod/2.f);

    gl_Position = pos;
}

my config in my gfx.c

// #define USE_GRAPHICS_PIPELINE
// #define NSIGHT
// #define RENDERDOC
// #define VERBOSE_VALIDATION
#define FIF 3
#define MIN_IMAGES 3

// #define USE_SLANG
#ifdef USE_SLANG
const char *vert_path = "..\\slang\\output\\vert.spv";
const char *frag_path = "..\\slang\\output\\frag.spv";
#else
const char *vert_path = "..\\vcc\\output\\vert.spv";
const char *frag_path = "..\\vcc\\output\\frag.spv";
#endif

[cloud rivet]

https://clang.llvm.org/docs/LanguageExtensions.html#matrix-types

I think it's time to make the switch to clang

I saw gobium point it out in a shady issue thread

the matrix types

Currently, the element type of a matrix is only permitted to be one of the following types:

an integer type (as in C23 6.2.5p22), but excluding enumerated types, bool, and _BitInt types whose width is not a power of 2;

the standard floating types float or double;

a half-precision floating point type, if one is supported on the target.

also

   // Matrix-vector multiplication (builtin operation)
    vec3_t result = m * v;

I won't miss operator overloading in C now

:|

I need to be using this in both projects

I was thinking maybe dumping the from scratch in one hour thing, since I think it's getting in the way of the interesting parts

it was an interesting idea, but in practice, I'm not sure

also I learned more about vulkan today, specifically the difference between the semaphore used in VkAcquireImageKHR and the one used for presenting, and how I only need frames in flight count semaphores for the former and swapchain images count for the latter and why

plus all the graphics pipeline stuff I had never done before

which is not really all that much different than SO tbh

just a bit more struct spam

I think I'll add the software renderer to palinode as an alternative renderer

and then I can reuse the same C code in the shader

and dump the silly from scratch in one hour, but just keep doing the video recordings

[cloud rivet]

hrm I wonder how using clang will impact raddbg though

[astral hinge]

hopefully not at all

[cloud rivet]

I'm on Ryan Fleury's discord he's complained about clang debug info a little bit 😅

I was using raddbg with zig though with no problem and that was at the time using a llvm backend

the clang version that ships with MS is a little older than what the clang docs have

I'm going to see if I can download that instead

4GB 😅

[cloud rivet]

clang builds noticeably slower

msvc is instant

clang hangs a second

idk maybe not!

[echo crystal]

you could turn it in software rasteriser series

[cloud rivet]

Yes

[cloud rivet]

I think I will just take the compile time hit and go with clang since that’s what the shaders use. I wonder if I can tell vcc where my clang is so it pulls in the same version and not the Visual Studio version. I will take a look later

I get matrices for free now

I wonder if I can get things like the adjugate of a matrix with it

Adjoint or whatever

I will probably still need some manual matrix math

It also has vectors which I am already using in the shader

[cloud rivet]

vcc wraps clang, I need to understand how vcc works a bit better

Exposing vulkan and spirv as language extensions is brilliant and feels nice

That’s how I can likely get non uniform indexing

[bronze socket]

definitely share your findings

playing with vcc is one of those things that have been in my backlog for One Day™ forever

[cloud rivet]

it's already been amazing for learning

I didn't know anything about clang or language extensions, so looking into matrices and adding the math operations like fmodf to my shader was an aha moment

that makes the code not portable but also awesome at the same time

I just typed s/less/not/ and Discord updated my previous message

[bronze socket]

wdym not portable

[vagrant musk]

yup, that has been a feature for a while

[cloud rivet]

I think clang's language extensions are not portable to other compilers

like I'm stuck with compiling my application with clang if I use them

but I get the benefit of sharing code across my shaders and all the features

[bronze socket]

is it the kind of stuff you can't write a native equivalent of?

like write some kind of header where you block off the vcc version and the cpu version with preprocessors

[cloud rivet]

yes they are language extension, they add syntax that are not valid standard C

A * B for example for a mat4_t type

there's no way way write that with C without the language extension

oh

I see what you're saying

that feels rough

[bronze socket]

it's pretty much standard procedure when sharing CPU-GPU code

hell it's pretty standard for C/C++ everywhere

how do you think libraries are cross platform

[cloud rivet]

well I don't have to do any of that if I use clang

the non-clang version of the math would explode the LOC where it could all be on one line with clang

idk

ok I will correct: still possibly portable, if I am not lazy 😅

[dark saffron]

gcc does have the vector extension at least

[bronze socket]

do you use a swizzle ext? that's the main thing I can think of that might be a definite no go

[dark saffron]

it's part of Clang's opencl support stuff

but you can also use nasl instead, or glm or whatever you want

https://github.com/shady-gang/nasl/

nasl has swizzles implemented with C++ fuckery

the native types are just there to do i/o

for our path tracer sample application, we use nasl on the host too, it's just plain C++

[bronze socket]

oh neat, looking at the swizzling code now

I never really thought it was possible without a compiler ext

[cloud rivet]

yes it has swizzle!

it's amazing

[dark saffron]

that was a one-day C++ hacking adventure 🐸

from a guy who doesn't really care about C++

[cloud rivet]

I was like can I do .x ?, and then it worked! vcc ❤️

[bronze socket]

btw while you're here, what are your thoughts about the builtin cuda target for clang

[dark saffron]

i have no thoughts ig

[cloud rivet]

I won't get swizzling with nasl since I use C only

I think

[dark saffron]

you'll get it with the Clang language extensions and the native vectors, but you'll have to use clang on the host too

[bronze socket]

rip, I was thinking you might've investigated it as part of your research, I only found out it existed recently so I was wondering how it held up considering that hypothetically it'd encounter the same issues as you did

[cloud rivet]

yes, I will use clang

[dark saffron]

what ? no, afaiu it only targets PTX

just an exercise in nvidia mainlining their needs

again afaiu

[cloud rivet]

I noticed those cuda extensions yesterday after looking through the language extensions, was surprised, it I got the sense of clang scope creep, but I don't know

[bronze socket]

yeah ptx is what I meant, but didn't you have issues with needing to represent structured control flow to the compiler

my main curiosity is the codegen quality

[dark saffron]

nvidia doesn't use structured control flow

they have explicit masks used in their intrinsics

[bronze socket]

I see, so they shouldn't have any issues generating equivalent quality code to vcc

or vice versa

[dark saffron]

it's such a different model a comparison is hardly useful

and these things aren't really to do with code quality, they're more correctness at this point

being able to carefully construct programs that take advantage of insight into GPU topology

but yeah there is an open problem here with vcc, and i'm waiting for clang to ship their new reconvergence tokens

shady is ready for maximal reconvergence otherwise

[bronze socket]

yeah that's what I meant by quality, initially I was thinking there was a chance they'd generate "CPU-y" code, which in my mind is low quality, assuming it produces the correct result otherwise

[cloud rivet]

those are in 21?

[dark saffron]

idk

[dark saffron]

what is cpu-y code

[bronze socket]

code generated with the assumption that it's running single threaded on a CPU

[cloud rivet]

https://clang.llvm.org/docs/AttributeReference.html#convergent

hrm

this may be different

[dark saffron]

there's no real difference when looking at IR, it's all scalar at the levels I or CUDA is working at

with some intrinsics thrown in

scheduling branches, vector/scalar regalloc and masking codegen is the responsibility of device-specific JIT compilers in drivers

[bronze socket]

I see, so you don't really do any major transformations to the code to make it more "GPU aware", you just make all those annotations available to the user and make sure everything's plumbed through?

e.g. what's uniform/non uniform, where barriers are inserted, etc.

I'm just trying to get a rough feel for what I should expect a shader/kernel compiler to be responsible for that runs the risk of surprise shitty codegen

but it might not be meaningful, because I don't think glslang or dxc do anything fancy to your shader code beyond translating your intention to spirv, right?

[cloud rivet]

well if there is some way to write some poorly performing code in the context of this project and vcc I am sure to find it

[dark saffron]

hm yeah kinda

[dark saffron]

some of the most important performance aspects are decided by the downstream compiler, so if you upset it by generating something it doesn't deal with well, you can hurt 🅱️erf

e.g. register pressure it can't optimize away

[dark saffron]

neither does vcc, conceptually

shady is complicated for completely annoying, bad legacy reasons

that's part of what made publishing difficult

a lot of the work that went in it is overcoming "artificial" obstacles we as an industry made for ourselves

by separating out gfx/compoot and never really moving away from the crappy self-limiting mindset of early SLs

large parts of SPIR-V for Vulkan just restate glsl limitations in an ssa ir formalism

[bronze socket]

that makes sense, good to hear that it's "only" legacy crust and not something that requires a bunch of additional compiler optimization work just to hit par with existing tools

definitely makes adoption way less risky

[cloud rivet]

hrm, probably some risks since you can do things that aren't native to the ir like function pointers and generic pointers

but if you don't use those those things then you can hit par, but then that is kind of defeating the purpose of vcc and using a real language

[bronze socket]

there's way more reasons to not use glsl/hlsl even without those features

full power C++ TMP for one

[cloud rivet]

slang's generics feels nicer than C++ templates fwiw https://shader-slang.org/slang/user-guide/interfaces-generics.html

I haven't used templates much in C++, only a little though

probably not comparable to "full power C++ TMP"

[cloud rivet]

Didn’t do a speedrun training thing today and felt much better and fun.

[cloud rivet]

I tried my hand at the Clang language extensions ext_vector_type and matrix_type and I had unrealistic expectations of how all that would work, but I like them.

I expected to be able to have some kind of nice way to write a matrix literal, but the initialization for a matrix type is a TODO in the draft so I made a helper function for readable matrix literals, for that so that's no problem, I think maybe the compiler will be kind with that and optimize it out so there's no run time cost, I could check godbold, but I don't really actually care right now.

I expected to be able to multiply a matrix with a vector (lol) and that's a no go, but it's trivial to convert a vector to a column matrix with __builtin_matrix_column_major_load.

the matrix type is definitely a WIP, but this is a for fun project, crazy stuff goes, I'm gonna make it work. I really need to read all the fine details of the language extension details. I was like learning on the go while recording and that was hard.

I'll have to write an inverse function and start building out a math library. I can't actually assume anything about internal memory layout since the draft says that is implementation defined, so I am not sure what I can assume about how this would work when I send data from host to the GPU and use the matrix_type in both the application and in my vcc compiled shaders.

I have to still carefully read the language extension docs to a complete picture and avoid mistakes and problems.

being able to do A * B is really nice and A * toMat(v) is also ok, and being able to swizzle vectors is gold

[cloud rivet]

if I wrote my own math library without these language extension types I think it would be a worse experience

[cloud rivet]

I don't want to make videos anymore. It's dumb

I just like programming

[dark saffron]

this is one for the pipeline

[cloud rivet]

I'm just trying to have fun and share what I'm doing but it comes across as quasi educational and nobody should take any words of mine as something to learn from

[cloud rivet]

I think I need this https://github.com/stevearc/conform.nvim

[echo crystal]

"learning software rasterising day1"

[cloud rivet]

yeah I hedged and provided lots of disclaimers, but anyway I feel nothing but relief now

[echo crystal]

that's good

[cloud rivet]

this focusrite thing has been amazing though, my audio out, not just my mic is so much better

[broken fog]

it's good ye

i have it set to format on save

[cloud rivet]

yes, that's what I will do too

[broken fog]

can't live without an autoformatter tbh

[cloud rivet]

having to indent everything manually felt like I was using notepad.exe

[broken fog]

or xcode

[cloud rivet]

I mean, I know how to do block selection

[broken fog]

yea but not just indents, honestly i just type complete garbage and let the autoformatter handle it

[cloud rivet]

same

[cloud rivet]

a cpu software rasterized triangle blitted to the HWND device context using the clang opencl vector language extension types

ok going to figure out how to do that on the gpu via vulkan compute dispatch

I need to figure out how to write vulkan gpu memory allocation code that I previously just used VMA for

I think I should be able to reuse some of this code with vcc

[cloud rivet]

I was thinking after I do the gpu raster, I would do a RT version too

then I'd have a triangle via cpu rasterization, gpu compute, graphics pipeline and RT, and then maybe I just work my way up in parallel across those, idk

RT via vulkan apis, actually don't know how good my gpu is for that

it's kind of dated

NVIDIA GeForce RTX 3070

[alpine basin]

It's good

Any RTX GPU supports Vulkan ray tracing

[cloud rivet]

added interpolation to my software renderer

I love clang's vector type

I didn't even have to look it up, just used the ol brain and my understanding of barycentric coordinates

and of course I knew to convert to linear and back to srgb

it's actually a mix, not interpolation

well, I mean I have a lerp I wrote, and a mix and I used the mix, anyway

fun

it's kind of cool, this is all me. this isn't me passing vertices through a vertex shader and it doing the math for me, this is all my code, rasterizing and rendering. I wrote all of those pixels

the background colors oscillate that's why they look different from before

[cloud rivet]

that's kind of an interesting artifact along the edges

I have no idea what that is

it's the video

it's not in the actual render

hrm it is

[broken fog]

looks like aliasing

[cloud rivet]

Yeah

I like C it is good

Feels similar to when I wrote zig

[cloud rivet]

Once I have all the triangles I will start on an imgui

Just some text, text input and a slider to start

ASCII only

[cloud rivet]

Once I have image creation working with vk I will blit my software raster bitmap to vk instead of the win32 HWND

then I can build the imgui once and it will work regardless of how anything rendered

[broken fog]

honestly yeah it's pretty nice to write

except when it isn't, there are times i do miss having operator overloading and generics

but the simplicity is good

[cloud rivet]

the clang vector & matrix types let me use math operators at least

but they are draft stage and clang specific language extensions

there's a vector type language extension that is compatible with a gcc extension though, I'm not using that one

alright time to start on creating images and buffers in vk, I have always referred to VMA docs I don't think I ever read much of the vk spec on memory I will start with that

just two chapters, 11 & 12

[cloud rivet]

ah there's an appendix for memory also

[cloud rivet]

references to Google's Fuchsia OS in the vk spec heh

[cloud rivet]

I'm just going to start by creating a staging buffer and a device buffer pushing a small value to it and trying to read it from my shader and see what happens

using renderdoc to debug what's on the device and visible to the shader

the original vk 1.1 vulkan tutorial has some example code too

looks like you create a staging and device buffer, allocate some memory for a staging buffer and a device local buffer, bind the memory for each of the buffers, and then map memory to staging memory and then copy memory from staging buffer to device buffer

[cloud rivet]

it doesn't look like much but it's just a proof of concept I got working, Pushed this triangle to the right with a float value I put into a buffer which is made known to the shader via push constant bda

so now I can allocate gpu memory, create vk buffers, map data to host visible buffers, copy data from my staging buffer to the gpu memory, bind memory, create a BDA and use it in a vcc compiled shader

only thing left is creating an image, I will work on that tomorrow

then I'll blit software rasterized bits to vk, and then I'll work on the compute software rasteriser

[cloud rivet]

it's kind of lame how my render is just a triangle, and will be for a bit as I put in more scaffolding. It'll get start looking better though soon enough

I want to figure out the compute rasterization and the RT stuff

then I'll work on more interesting looking problems

[cloud rivet]

I have a hard time understanding some aspects of vulkan sync, and haven’t gotten to the point where I intuitively understand when there’s a hazard

I have this execution dependency I resolved with an extra barrier that exists only because I get VVL sync errors telling me I need it. The message is bizarre

It said I needed a VK_PIPELINE_STAGE_2_NONE as a destination stage

I resolved it just with an extra barrier that just matched the stages I was actually using and I already had a barrier for

It seemed to say that that the memory barrier that transitioned the swap image after acquiring it was a write after read hazard

why would another barrier solve that?

I guess maybe because it is just an execution barrier it ensures the write doesn’t start until the read from acquire is done. I should read about acquire more

Well

That explains it

TIL

I guess I just accidentally got this right in my previous project

[vagrant musk]

Vulkan sync

[cloud rivet]

The vulkan tutorial doesn’t have this barrier

It just has two

Hrm

Oh that makes sense though

It just presents

Ok ok

I can't stand it when I don't understand why something in my code works, I should be able to understand why I need every barrier I have, and now I do and the world makes sense

I just had to write it out

it's not reading from the acquire though, it's the presentation engine that is doing the reading

I guess assumed that the acquire handled all of this, like the image is officially "available" but I guess it isn't, not really. Acquire next available image that you still additionally have to make sure is actually available

got it

[cloud rivet]

ahh, I like put the diff of the before and after with the sync and the operations I am making side by side in two text editors and I get it now

what I need is a way to print this out as debug info

so I can see it just like this

clearly I was missing a barrier

vkAcquireNextImageKHR
vkBeginCommandBuffer  

image_barrier(cf.vk_cmd_buffer,
                swapchain_image,
                VK_IMAGE_LAYOUT_UNDEFINED,
                VK_IMAGE_LAYOUT_GENERAL,
                VK_IMAGE_ASPECT_COLOR_BIT,
                VK_PIPELINE_STAGE_2_COLOR_ATTACHMENT_OUTPUT_BIT | VK_PIPELINE_STAGE_NONE,
                VK_PIPELINE_STAGE_2_CLEAR_BIT,
                VK_ACCESS_2_COLOR_ATTACHMENT_WRITE_BIT,
                VK_ACCESS_2_TRANSFER_WRITE_BIT);
  {

vkCmdClearColorImage
vkCmdBeginRendering

// bind SO, configure graphics pipeline, push constant, draw triangle

vkCmdEndRendering

image_barrier(cf.vk_cmd_buffer,
                swapchain_image,
                VK_IMAGE_LAYOUT_GENERAL,
                VK_IMAGE_LAYOUT_PRESENT_SRC_KHR,
                VK_IMAGE_ASPECT_COLOR_BIT,
                VK_PIPELINE_STAGE_2_CLEAR_BIT,
                VK_PIPELINE_STAGE_2_COLOR_ATTACHMENT_OUTPUT_BIT,
                VK_ACCESS_2_TRANSFER_WRITE_BIT,
                VK_ACCESS_2_NONE);

vkEndCommandBuffer
vkQueueSubmit2
vkQueuePresentKHR

and now it's

vkAcquireNextImageKHR
vkBeginCommandBuffer

image_barrier(cf.vk_cmd_buffer,
                  swapchain_image,
                  VK_IMAGE_LAYOUT_UNDEFINED,
                  VK_IMAGE_LAYOUT_GENERAL,
                  VK_IMAGE_ASPECT_COLOR_BIT,
                  VK_PIPELINE_STAGE_2_COLOR_ATTACHMENT_OUTPUT_BIT,
                  VK_PIPELINE_STAGE_2_CLEAR_BIT,
                  0,                             // src access
                  VK_ACCESS_2_TRANSFER_WRITE_BIT // dest access
);

vkCmdClearColorImage

image_barrier(cf.vk_cmd_buffer,
                swapchain_image,
                VK_IMAGE_LAYOUT_GENERAL,
                VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL,
                VK_IMAGE_ASPECT_COLOR_BIT,
                VK_PIPELINE_STAGE_2_CLEAR_BIT,                   // src stage
                VK_PIPELINE_STAGE_2_COLOR_ATTACHMENT_OUTPUT_BIT, // dst stage
                VK_ACCESS_2_TRANSFER_WRITE_BIT,                  // src access
                VK_ACCESS_2_COLOR_ATTACHMENT_READ_BIT            // dest access
  );

vkCmdClearColorImage

vkCmdBeginRendering

// bind SO, configure graphics pipeline, push constant, draw triangle

vkCmdEndRendering

image_barrier(cf.vk_cmd_buffer,
                swapchain_image,
                VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL,
                VK_IMAGE_LAYOUT_PRESENT_SRC_KHR,
                VK_IMAGE_ASPECT_COLOR_BIT,
                VK_PIPELINE_STAGE_2_COLOR_ATTACHMENT_OUTPUT_BIT, // src stage
                VK_PIPELINE_STAGE_2_COLOR_ATTACHMENT_OUTPUT_BIT,
                VK_ACCESS_2_COLOR_ATTACHMENT_WRITE_BIT, // src access
                0);

vkEndCommandBuffer
vkQueueSubmit2
vkQueuePresentKHR

I just need a way to print this ^^

I'm going to add a way to debug print my render graph

just so obvious it was wrong

I thought I had cleaned up my sync, maybe I accidentally deleted a barrier when I added the graphics pipeline code

the or with 0 is also funny

| VK_PIPELINE_STAGE_NONE lol

jfc

[cloud rivet]

I'm just going to use a linked list for my first iteration of my render graph tbh

using my per frame arena, just recreate it every frame for now, just to get something going

and then interate

this doesn't actually allocate any new memory, since the arena allocation happens at application start and I just reset it every frame

this is better because it is declarative, I don't have all the vk struct spam and api calls to scroll through to see what I'm doing

I'm not sure, will see

[cloud rivet]

just sort of outlining how it might work in code this isn't functional yet

[cloud rivet]

it works :D

no sync vvls or anything

now extending my render to do more complex things is just extending the graph

and adding additional graph node types

and handlers for them

way better than a giant render function where image barriers are 5 pages of code apart 😅

adding a blit for my software rasterizer tomorrow, then add a draw image and then add a software rasterizer in compute, then an RT triangle and then start on 3D stuff

[cloud rivet]

I am going to move the graphics pipeline to use a mesh shader

I am on the ultimate single triangle render bike shedding journey

This is an enterprise triangle

[broken fog]

fake, not written in java

[vagrant musk]

Enterprise java makes me want to drive a boat off a bridge

[cloud rivet]

True

Time to rewrite in Java jk

I will use uniscribe for text

Shameless win32 only

Wine supports uniscribe

Maybe I try and get my project to work on steamdeck with proton at some point

Now I am no longer ascii only

[cloud rivet]

I wish I had known that you can put a break point in the validation debug handler a year ago

that is so incredibly convenient

the debug handler is synchronously fired where the error happens and it is in the call stack

so you immediately know what is wrong

[cloud rivet]

I really feel like I hit a stride with my vulkan sync now. I added a per fif draw image and clear and render to it now and then blit it to the swapchain image, wrote out how the image barriers worked and got it mostly right. I did get the read and then write access masks wrong for the barriers that sandwich the dynamic rendering but I understood the issue, I never felt lost, was easy to fix

it's so easy to see what's going on now too

  // Wait for swapchain image present
  render_node_cfg_t acquire_barrier_cfg = {0};
  acquire_barrier_cfg.image_barrier_cfg = (image_barrier_cfg_t){0};
  acquire_barrier_cfg.image_barrier_cfg.old_layout = VK_IMAGE_LAYOUT_UNDEFINED;
  acquire_barrier_cfg.image_barrier_cfg.new_layout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
  acquire_barrier_cfg.image_barrier_cfg.aspect_mask = VK_IMAGE_ASPECT_COLOR_BIT;
  acquire_barrier_cfg.image_barrier_cfg.src_stage_mask = VK_PIPELINE_STAGE_2_COLOR_ATTACHMENT_OUTPUT_BIT;
  acquire_barrier_cfg.image_barrier_cfg.src_access_mask = VK_ACCESS_2_NONE;
  acquire_barrier_cfg.image_barrier_cfg.dst_stage_mask = VK_PIPELINE_STAGE_2_BLIT_BIT;
  acquire_barrier_cfg.image_barrier_cfg.dst_access_mask = VK_ACCESS_2_TRANSFER_WRITE_BIT;
  acquire_barrier_cfg.image_barrier_cfg.image_type = image_barrier_type_swapchain_image;
  current_node = add_render_node(actx,
                                 arena,
                                 current_node,
                                 "wait for present read on swapchain image before blitting to it",
                                 render_node_type_image_barrier,
                                 &acquire_barrier_cfg);
  // Prepare draw image
  render_node_cfg_t reset_draw_barrier_cfg = {0};
  reset_draw_barrier_cfg.image_barrier_cfg = (image_barrier_cfg_t){0};
  reset_draw_barrier_cfg.image_barrier_cfg.old_layout = VK_IMAGE_LAYOUT_UNDEFINED;
  reset_draw_barrier_cfg.image_barrier_cfg.new_layout = VK_IMAGE_LAYOUT_GENERAL;
  reset_draw_barrier_cfg.image_barrier_cfg.aspect_mask = VK_IMAGE_ASPECT_COLOR_BIT;
  reset_draw_barrier_cfg.image_barrier_cfg.src_stage_mask = VK_PIPELINE_STAGE_2_BLIT_BIT;
  reset_draw_barrier_cfg.image_barrier_cfg.src_access_mask = VK_ACCESS_2_NONE;
  reset_draw_barrier_cfg.image_barrier_cfg.dst_stage_mask = VK_PIPELINE_STAGE_2_CLEAR_BIT;
  reset_draw_barrier_cfg.image_barrier_cfg.dst_access_mask = VK_ACCESS_2_TRANSFER_WRITE_BIT;
  reset_draw_barrier_cfg.image_barrier_cfg.image_type = image_barrier_type_draw_image;
  current_node = add_render_node(actx,
                                 arena,
                                 current_node,
                                 "reset draw image before clearing the color",
                                 render_node_type_image_barrier,
                                 &reset_draw_barrier_cfg);

  current_node = add_render_node(actx, arena, current_node, "clear color", render_node_type_clear_color, NULL);

  render_node_cfg_t clear_barrier_cfg = {0};
  clear_barrier_cfg.image_barrier_cfg = (image_barrier_cfg_t){0};
  clear_barrier_cfg.image_barrier_cfg.old_layout = VK_IMAGE_LAYOUT_GENERAL;
  clear_barrier_cfg.image_barrier_cfg.new_layout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL;
  clear_barrier_cfg.image_barrier_cfg.aspect_mask = VK_IMAGE_ASPECT_COLOR_BIT;
  clear_barrier_cfg.image_barrier_cfg.src_stage_mask = VK_PIPELINE_STAGE_2_CLEAR_BIT;
  clear_barrier_cfg.image_barrier_cfg.src_access_mask = VK_ACCESS_2_TRANSFER_WRITE_BIT;
  clear_barrier_cfg.image_barrier_cfg.dst_stage_mask = VK_PIPELINE_STAGE_2_COLOR_ATTACHMENT_OUTPUT_BIT;
  clear_barrier_cfg.image_barrier_cfg.dst_access_mask = VK_ACCESS_2_COLOR_ATTACHMENT_READ_BIT;
  clear_barrier_cfg.image_barrier_cfg.image_type = image_barrier_type_draw_image;
  current_node = add_render_node(actx,
                                 arena,
                                 current_node,
                                 "wait for clear color write before rendering to draw image",
                                 render_node_type_image_barrier,
                                 &clear_barrier_cfg);
  current_node
      = add_render_node(actx, arena, current_node, "begin rendering triangle", render_node_type_begin_rendering, NULL);

now that I have a draw image per fif I can take my software raster bits and upload them via my staging buffer and then copy the bytes to the draw image

and after that I'll start on the gpu raster which will also write to the draw image when it's being used

I'm not sure how RT will work I don't know anything about RT yet

it may seem a bit silly but what I basically end up with is a way to draw from all sorts of places, I could combine them all into one present, I'm not going to be limited to just drawing from the graphics pipeline anymore

and it just ends up being all about how I set up and configured the render graph

[cloud rivet]

Claude says that copying my bitmap to my staging buffer to copy to my draw image which is blitted to the swapchain image would be more efficient than using StretchDiBlit because I am bypassing the window stack. I guess I will find out

I will keep StretchDiBlit around to measure

That entire bitmap has to go to the GPU every frame anyway since that’s what my monitor is plugged into. Why would me manually copying to the GPU be slower

It’s gotta be comparable

If it’s faster that’s even better

Just gotta measure

[cloud rivet]

I will need per fif staging buffers

Or regions

I could rasterize in threads and always grab the latest finished one and copy it over

[bronze socket]

sounds like you got some exceptionally weird advice

StretchDiBlit is GDI software blitting, as in on your CPU

I think you either got a hallucination or a regurgitation of weird reddit misinfo

[cloud rivet]

No I didn’t provide the full context here. I asked about a comparison of the entire process

A comparison of using StretchDiBlit to the HWND device context to just copying that same bitmap via vk to my swapchain

Like end to end either way it has to get to my screen

Through the GPU

Right

[bronze socket]

also don't forget there's a compositing step

[cloud rivet]

I expect both to take roughly the same amount of time

What is the composting step?

[bronze socket]

where your OS blits the swapchain image onto the actual image of your entire monitor

alongside all the other windows

[cloud rivet]

The bottleneck has to be the massive data transfer though right

Like any of this other stuff probably doesn’t matter?

[bronze socket]

PCIe is wicked fast

the problem I am seeing here is what timeline you're burning

if you're using StretchDiBlit you're burning CPU time on a ram-to-ram write

that will have to be uploaded anyway

[cloud rivet]

I just don’t want my vk process to be slower than using SterchDiBlit to the dc and whatever that does

It sounds like you’re agreeing with claude

[bronze socket]

yeah I thought it told you the opposite

[cloud rivet]

So me doing this through vk is actually a good plan

[bronze socket]

also if you really want to needlessly minmax this you can look into stuff like your GPU's transfer queue

[cloud rivet]

Oh right

Those are faster yes

[bronze socket]

and issuing a copy command vs writing into host visible memory

[cloud rivet]

Hemm

[bronze socket]

not faster, they keep it off your main gpu timeline

[cloud rivet]

Oh right

[bronze socket]

so you can transfer while doing other work

[cloud rivet]

Exciting

[bronze socket]

also to fill your head with more pointless things to try to optimize something that barely matters - vblanco reported that you can get surprisingly faster results reading from device-visible host memory in a compute shader and doign your "copy" that way

albeit that was more in the context of buffers

for images I don't think it's necessarily the same story

[cloud rivet]

Nice

I plan on also adding a GPU rasterizer so I can just repurpose it

For the cpu software rasterized bits

It’s just rendering a triangle but this renderer is already so much better than what I had in Rosy for so many reasons

The only that thing has that this doesn’t is MSAA which is trivial to add

Also it has mipmaps, I need to add that

No point to do that until I write a png decoder though

The png,jpeg and JSON code (for GLTF) will be the least fun code I will write I hope

Maybe I need exr and compression code also though

[bronze socket]

if it's not fun why not just use stb image

and fastgltf

[true moon]

I would just only support uncompressed tiff personally if I were forcing myself to go without libs

[bronze socket]

and good ol' bmp

[cloud rivet]

Nah, Id just use Windows Imaging Component if I wasn’t doing everything from scratch

[bronze socket]

I wrote my own readers for ktx and dds now that I think about it

it's just headers + binary data

[cloud rivet]

dds is super easy

The compression is hard

[bronze socket]

ktx is even easier because the headers are slightly less dumb

and working from khronos specs is nice

[cloud rivet]

I am going to learn a lot from doing all this from scratch

I am looking forward to the serotonin I get once I see a texture render that I decoded myself

It’s just the process will be the least fun since it all has to work before I can see anything

work end to end

The only not from scratch is the text, it’s just way too hard

I will call out to windows for it

[bronze socket]

yeah that's probably fair, cleartype is very nice as well

[cloud rivet]

Even handmade hero code uses windows for fonts

[cloud rivet]

man my structs are a crime against computers, the laws of nature, and good programming

https://github.com/EpicGamesExt/raddebugger/blob/master/src/raddbg/raddbg_core.h#L540

yes team fat struct

I am not alone

I see similar structs in hmh code

[cloud rivet]

oh I figured out some new C stuff

I have a common.h

which had a bunch of my majorly big structs

I wanted to separate out their definitions into files because they were sort of piling up in common.h

and I am using C style typedefs since I am writing C

and wasn't sure how to forward declare those

but I figured it out, I just typedef them where I forward declare them and then actually define the struct elsewhere

struct DescriptorSetAllocator;
typedef struct DescriptorSetAllocator DescriptorSetAllocator;

struct GraphicsContext;
typedef struct GraphicsContext GraphicsContext;

struct AppContext;
typedef struct AppContext AppContext;

struct CPURasterContext;
typedef struct CPURasterContext CPURasterContext;

so I reduced the struct spam in my common.h

I love C, it is fun

I haven't missed the std lib at all or raii, I just have a couple of arenas and I never worry about memory ever

I allocate 512MB of memory in two arenas, way more than I currently need, and nobody came and arrested me and my app is fine

[astral hinge]

I wonder why they have various aliases for static

[cloud rivet]

I do this too

static can mean multiple things

[astral hinge]

I guess it communicates intent, but I didn't know what it meant at first

[cloud rivet]

it's just to make it more readable yeah

[astral hinge]

also they call const read_only

oh wait

it's something else entirely

[cloud rivet]

sorry I thought you were referring to internal

[astral hinge]

it puts it into rodata

[astral hinge]

that and global

[cloud rivet]

yeah

same has hmh

I learned it from casey muratori

[astral hinge]

and local_persist

[cloud rivet]

I think these are Rad Game Tools idioms

all the people who have worked at Rad Game Tools write very similar C

and now they're all like super influencers in low level and game dev programming communities

so it has spread

[astral hinge]

programming influencers 😩

[cloud rivet]

I mean

it's true though

super influential

[astral hinge]

they are

[cloud rivet]

but in a good way?

[astral hinge]

they're on the opposite end of the spectrum as those hardcore OOPers like uncle bob

[cloud rivet]

yes

❯ Measure-Command { cmd /c 'build.bat' }


Days              : 0
Hours             : 0
Minutes           : 0
Seconds           : 0
Milliseconds      : 850
Ticks             : 8506065
TotalDays         : 9.84498263888889E-06
TotalHours        : 0.000236279583333333
TotalMinutes      : 0.014176775
TotalSeconds      : 0.8506065
TotalMilliseconds : 850.6065

my current build time from scratch

just 4k lines of code though

I think I can just have it print out the milliseconds

❯ Measure-Command { cmd /c 'build.bat' } | Format-List -Property Milliseconds


Milliseconds : 790

idk

I actually don't know who uncle bob is

[astral hinge]

"clean code" guy

[cloud rivet]

I have two copies of that book

one for me and one I wanted to give to someone else so they would learn what was in it

I have learned since then it was all really not good advice

but when I first read I was all into it, I don't know like 10 or 15 years ago

ah

I didn't know he had that nickname

[astral hinge]

he's not really my uncle in case you were wondering

[cloud rivet]

I got a jury summons

I haven't ever been called before

but I gotta go monday

:\

I hope I can get out of it

maybe they see my long hair and go, nope not that guy

[astral hinge]

there will be a bunch of candidates and they (the lawyers) will ask you all questions to eliminate jurors that might be biased

so there's that, or you can just whisper "jury nullification" into one of the attorney's ears

[cloud rivet]

lmao

Judges hand out contempt of court charges like giving out candy on halloween. I am going to behave I think

[astral hinge]

but then you could brag to everyone about your rap sheet

[cloud rivet]

I would lose my job lol

Can’t have a criminal record

It’s a security company

[astral hinge]

just keep the bragging to your close friends then

I'm jk. Good luck out there

I like to frame jury duty as a novel diversion rather than a boring thing you're being forced to do

Reframing is a powerful tool

[cloud rivet]

yes

[true moon]

Your odds of actually being chosen are very low

Generally you just go sit in a waiting room and then they tell everyone the defendant pled guilty once they saw the jury candidates coming in and the reality of the situation hits them

Either that or they select their jury and everyone else goes home

And in the worst case if you actually get picked then you have paid leave to step out of the workforce and do your civic duty which is neat

Also by living in Alameda county you have the distinct advantage of being close to your courthouse, for some reason in Santa Clara county the courthouse me and all my family members have been summoned to is in Morgan Hill 💀

[bronze socket]

the lawyers give you really easy outs

the question I got was "would you take a cop's testimony over a civilian's" and I just said yes and that was that

[bronze socket]

you can go online and submit a request to change the location

[wraith urchin]

Is it weird that I kinda want to do jury duty?

[bronze socket]

have you seen 12 angry men

mandatory viewing

[wraith urchin]

I actually haven't lol

But I just think its like actually pretty cool that the average person has the opportunity to participate in legal proceedings

Like if they don't pick you, someone else is gonna end up on the jury, do you trust that person?

I mean granted maybe things are a little different since I live in a country where judges still occasionally wear powdered wigs

[cloud rivet]

Yes I agree, unfortunately my work will suffer and I will come back to a bad mess. But I agree jury duty service is a good thing overall

[cloud rivet]

I would like to track images and buffers as stateful objects in my render graph and have their uses update the stages and masks automatically instead of still doing it declaratively

Hrm

I am not going to do that yet

Or maybe I should before I add more stuff

Should not be hard

[bronze socket]

there is a saying that in a jury trial, your life depends on 12 guys who were too dumb to get out of jury duty

I wouldn't wanna do it personally because it's pure tedium, but yeah it's absolutely wonderful you can have your fate decided by normal people and not a handful of lying sophists

but in practice the odds you'll be some kind of rare hero of justice in a blazing hot trial are extremely slim

[cloud rivet]

I also just prefer not being around people, especially people I don’t know

[true moon]

Me too generally but this is a rare opportunity to participate in government with a random assortment of your fellow citizens

Getting on an actual jury is a once in a lifetime experience

[true moon]

Yeah I was kinda disappointed when they dismissed us all

[wraith urchin]

I got a letter telling me I was on the rotation but I haven't been summoned yet

[cloud rivet]

I'm going to get this software rasterizer bitmap to vk thing going, and then I'll mess around with tracking image/buffer sync state. My brain wants to believe that I should be able to read the spec and learn from that what stage and access mask operations I do require and with tracking previous sync stage and access know what the next one should be based on the operation and that should just work and sync should be a mostly solved problem

and I won't even need to declare these things manually anymore

hrm

with buffers I think the offset or memory range of the buffer accessed is part of the resource that needs to be tracked, not just the buffer

anyway, gonna get this software raster copy thing working onw

a lot of engine bikeshed & fixed investment taking place before I get passed just a triangle

[cloud rivet]

the fuck

I use photoshop rarely, and yet somehow creative cloud uses more memory than anything else

how do I even exit this program

what's it even doing

ok

that did it

just disabled all of those

[cloud rivet]

so, I think maybe the first image format I implement is uncompressed exr

I was looking at OpenEXR and built the library and the tools and am able to take a polyhaven downloaded EXR and remove the compression from it with the openexr tools

C:\Users\Bjorn\Downloads>exrinfo charolettenbrunn_park_4k.exr
File 'charolettenbrunn_park_4k.exr':
  compression: 'piz'
  displayWindow: [ 0, 0 - 4095 2047 ] 4096 x 2048
  dataWindow: [ 0, 0 - 4095 2047 ] 4096 x 2048
  channels: 4 channels
   'A': float samp 1 1
   'B': float samp 1 1
   'G': float samp 1 1
   'R': float samp 1 1

C:\Users\Bjorn\Downloads>exrmaketiled.exe --help

C:\Users\Bjorn\Downloads>
C:\Users\Bjorn\Downloads>exrmaketiled.exe -z none "C:\Users\Bjorn\Downloads\charolettenbrunn_park_4k.exr" "C:\Users\Bjorn\Downloads\charolettenbrunn_park_4k_uncompressed.exr"

C:\Users\Bjorn\Downloads>exrinfo charolettenbrunn_park_4k_uncompressed.exr
File 'charolettenbrunn_park_4k_uncompressed.exr':
  type: 'tiledimage'  compression: 'none'
  tiles: size 64 x 64 level 0 (single image) round 0 (down)
  displayWindow: [ 0, 0 - 4095 2047 ] 4096 x 2048
  dataWindow: [ 0, 0 - 4095 2047 ] 4096 x 2048
  channels: 4 channels
   'A': float samp 1 1
   'B': float samp 1 1
   'G': float samp 1 1
   'R': float samp 1 1
  tiled image has levels: x 1 y 1
    x tile count: 64 (sz 4096)
    y tile count: 32 (sz 2048)

oh there's a cubemap option instead of tiled

I can actually just work my way from uncompressed to zip (which I can reuse for png) to piz (which uses huffman which I can reuse for jpeg)

exr is really well documented

https://openexr.com/en/latest/OpenEXRFileLayout.html

or maybe I go with png since there's working examples in hmh and stb_image, and stb_image also jpg

and then I try exr, exr looks trivial though

especially with no compression

idk I have other things to do before then

https://github.com/AcademySoftwareFoundation/openexr/blob/main/src/lib/OpenEXRCore/internal_piz.c

openxr's decompression code is all in C

and well documented

they have huffman and zip code too just as readable and documented

[cloud rivet]

noice

that's cpu raster to vk

that was more work than I anticipated, and, I got a bunch of validation errors, but I got it working, now have to clean it up

it's cool how blitting an image does format conversion

I actually need to stop making my blit map ARGB since I no longer have to play by win32 rules

a benefit I hadn't anticipated when I planned all this

[cloud rivet]

I have changed my mind on all this image & json stuff

I spent months on rsy asset format and that asset pipeline

I am just going to use it

It generates mipmaps and BC compression

[cloud rivet]

kind of cool being able to see my software raster in renderdoc now

obviously not helpful with debugging the render, but I can now see all the color data

there were a lot of unforseen benefits to changing it to sending it to screen via vk

[cloud rivet]

my bitmap is now using f16 half floats like 1.f16

typedef _Float16 f16;
typedef float f32;
typedef double f64;

typedef struct {
  f16 r;
  f16 g;
  f16 b;
  f16 a;
} vk_R16G16B16A16_sfloat;

  u64 *pixel = (u64 *)actx->crctx->bitmap_bits;

  {
    int quad_row = 0;
    for (int y = 0; y < actx->crctx->bitmap_height; y++) {
      for (int x = 0; x < actx->crctx->bitmap_width; x++) {
        vk_R16G16B16A16_sfloat color = {1.f16, 0.f16, 1.f16, 1.f16};
        memcpy(pixel, &color, sizeof(vk_R16G16B16A16_sfloat));
        pixel++;
      }
    }
  }

totally works

[cloud rivet]

am I spamming this thread too much

[astral hinge]

it's your thread fam

I read everything you post here even if I don't reply

[cloud rivet]

ok, I honestly don't know why I post so much, when I wasn't on discord I was trying to put all my thoughts in my notes app

and I never really did

although I have switched to obsidian

and it's much better

[astral hinge]

I view posting on discord as basically interactive notetaking

it's more painful to go back and read the "notes", but usually you don't need to do that anyway

[cloud rivet]

yeah, it's sort of like rubberduck debugging

[astral hinge]

the act of posting and discussing solidifies the ideas in your brain

[cloud rivet]

I'm going to work on tracking sync resource state so I don't have to declare it

then I will finally start on gpu software rasterization

last week was a lot of scaffolding work

[astral hinge]

how come you're doing gpu softrast

I didn't read

[cloud rivet]

just because it's interesting, and I think valuable maybe? because of vcc I should just be able to reuse the same code I think? or write the code to be reusable?

[astral hinge]

could be an excuse to implement some stuff you can't normally get with ff rasterization

but I'm out of ideas rn

[cloud rivet]

what's ff?

[astral hinge]

fixed function

like vkCmdDraw

[cloud rivet]

oh I see

yeah

I hope to find all the edge cases in vcc and annoy gob as much as possible with github issues

jk

[astral hinge]

I think gob would like that actually

[cloud rivet]

nah I just really am looking forward to writing a bunch C shader code

[astral hinge]

you will be a valuable test subject

[cloud rivet]

maybe at some point I can contribute a PR once I understand it better who knows

[astral hinge]

for vichichi

[astral hinge]

the code is clear enough that even I, someone with no background in compilers, could hack some changes into the glsl backend

[cloud rivet]

yeah it's all been pretty readable

[astral hinge]

for the spir-v backend you'd have to be familiar with spir-v though

[cloud rivet]

I have tried writing a little toy compiler once with yacc/bison one time

that was a long time ago

[astral hinge]

nice

[cloud rivet]

I think this will work for tracking image sync state

typedef enum {
  image_barrier_type_no_image,
  image_barrier_type_swapchain_image,
  image_barrier_type_draw_image,
} image_barrier_type_t;

typedef struct {
  VkImageLayout layout;
  VkImageAspectFlags aspect_mask;
  VkPipelineStageFlags2 stage_mask;
  VkAccessFlags2 access_mask;
  image_barrier_type_t image_type;
  size_t draw_image_index;
  size_t swapchain_image_index;
} vk_image_sync_state_t;

going to try this

I don't want to stick the actual VkImage on it

I'm not sure

just going to try this

hrm

actually I should stick this on my vk_image_t

typedef struct {
  VkDeviceMemory vk_memory;
  VkImage vk_image;
  VkImageView vk_image_view;
  VkExtent2D vk_extent;
  vk_image_sync_state_t sync_state;
} vk_image_t;

and remove those indicies from the sync state

I can actually now just keep track of it without changing my render graph and then transition to referring to its state as an iterative step

and see if that works, and if it does I can remove the explicit barriers and add the barriers on the render graph nodes that do operations that would add hazards

[cloud rivet]

searching through the spec to see how the sync required for operations I am doing is documented, for example for blit and copy it says

All copy commands are treated as “transfer” operations for the purposes of synchronization
barriers

I feel like when something operates on an image the spec should be really clear on these things

so when I do something that takes an image or a buffer I should know what layout it requires and what stage it happens in and what access masks are needed/impacted

that's not really sync, but that's the kind of details spec should also have for sync imo

[cloud rivet]

these things are just haphhazardly spread throughout the spec with no clear single place to figure it out

like sync is so important they created a validator for it

but then in the spec they just winged it

the "must be synchronized with" is just in chapter 38, 39 and chapter 4

I understand, the nature of which stages an operation occur in depend on how they've been configured

the Pipeline stage section in chapter 7 is the key then

sort of

[cloud rivet]

my image sync state render graph works yay

I pass in my vk_image_t to my image barrier helper function

it reads the old layout, stage mask and access mask, and then writes the new layout, stage mask and access mask and that works

no VVL sync errors

I do have to update the vk image access masks after render begin since that writes to the image given without an image barrier

and after acquiring a swapchain image I set its access mask to none

no VVL, no more hard coded image barriers where I manually have to hard code image states

it's all programmatically managed now

no ALL_COMMANDS

it's just a single queue render with a very simple render, I'm sure this gets harder

but this is so much better than any vulkan renderer I have written so far, which was just a 2000 line monolitic typed out render function in Rosy lol

I can finally work on gpu software rasterization and work on rendering stuff now

a two week vk scaffolding detour

[cloud rivet]

added keyboard input handling and can now switch between graphics pipeline render graph, shader object render graph and cpu raster render graph by pressing a number, going to add gpu raster next

the graphics pipeline graph and the shader object graph just render the exact same thing right now so you can't tell that's changing, but I confirmed in renderdoc

before it was command line arg

and that was annoying to test to make sure I wasn't getting vvls

[cloud rivet]

I used to think that every thing would take more time to write with C since there’s no stl and no generics but actually the opposite is true. Faster compile times, void *everything, arenas, no fucks given -> light speed development

Not like vblanco 2 weeks to make a Minecraft clone light speed

More like 3 weeks and still staring at a triangle light speed

[wraith urchin]

It took him like a year lmao, he just moved it to unreal in a few weeks

[cloud rivet]

Yeah that’s true

[cloud rivet]

Thinking about the uniform conversation in Grub’s thread I think I will actually try what that led me to think about. An OpenGL like uniform mechanism that lets me specify arbitrary data types I want available in my shader. So if I wanted to have a float4 or a float or a uint or mat4x4 I just call a function that creates this, gets an index back, and then I use that index to update the value and I send that index in with my push constant.

It will all be backed by a per fif buffer that is copied every frame. The index points into a table that returns the offset in bytes and type of the value. The bda will be into a global buffer that has all this available via the scene buffer. Maybe the just stick at the end of the scene buffer.

Hrm, this would be separate data from the scene graph that I get from an asset. Just arbitrary convenience data I could use when I wanted it

I like that

I don’t think I need a BDA in my push constants, they’ll just be global and the push constants wii have offsets

[cloud rivet]

I got excused from jury duty. Just silently sat in a chair for 1.5 days

Nobody even talked to me

[true moon]

Typical lol

[astral hinge]

civil service = accomplished

[cloud rivet]

I did my part

Since I am just going to use the rsy asset format I get the benefit from all my months of working on that and will just get mipmaps, mesh optimizations, mikktspace, bca compression. Only thing I need is copy over the Rosy scene graph and parse the dds header and have no problem loading sponza

No need to write a json parser or add image decoding

But I will have to write some C++ to extend it to get HDRI

So will do some work on Rosy

I guess that project is not dead yet

Not doing this work yet though

It’s all going to move real fast soon, once my software rasterizers and RT pipelines are working

[wraith urchin]

https://tenor.com/view/waking-up-anger-the-undertaker-wwe-gif-15072678

[cloud rivet]

Maybe eventually Palinode is just the renderer for Rosy

[vagrant musk]

Incredible

[cloud rivet]

I am going to try and add the image2D builtin type to vcc so I can write to an image in a compute shader. Not fully familiar with how the vcc IR works but this seems like a small task given I can follow the sampler builtins that are already present

[cloud rivet]

alright

been blocked on compute shader stuff because I want use descriptor indexing in my shader and that's not working due to vcc compiler errors

was assuming I was doing something wrong or it was maybe a small bug in vcc, which may still be the case, that it is a small bug

but without a broad understanding of how the compiler works my random attempts to be lucky with small changes that didn't understand how they would actually work failed, which course they would

so just gonna spend the weekend learning more

I have created a test case that fails in my fork of vcc

https://github.com/btipling/shady/pull/1/files#diff-a14b4d4ab4459569a94d0df9c8ac9cfec03a497be9534b1161247923638b4ab3

and while I haven't figured out the specifics I have pretty good understanding of how the frontend llvm works and how the backend spirv emit works

and I can read the IR dumps

I understand the address space a little bit

goal today is to understand how the descriptor annotation works right now without DI

like from C to spirv end to end

then I am just going to create my own node type in the shady IR and try to have it emit some simple spirv

so I understand that part

maybe as I grow to understand it I'll notice that this may be a small bug but just to get anything working at all with DI I want to actually bypass all the existing array address space stuff with a new annotation that just treats DI as a magical thing and to skip all the existing logic and just treats it as the opaque black box it is anyway

and then I can go back to my stuff

[cloud rivet]

ok ok I think I understand the annotations thing end to end, the LLVM IR representation, and how they're parsed into shady nodes and how the global variable nodes get and store annotations, and how they're referred to in the backend to emit spirv

I'm going to add some debug logging, step through it in the debugger and print out the LLVM IR and Shady IR and then I think I will feel pretty good about it

[cloud rivet]

https://llvm.org/doxygen/group__LLVMCCoreValueGeneral.html

there are cool

and helpful

like const char * LLVMGetValueName (LLVMValueRef Val) 🙏

[cloud rivet]

TEXSAMPLER: Converting LLVM global 'texSampler' at 0000022C979F0320
TEXSAMPLER: Created GlobalVariable node at 0000022C97D5DF50
Converting function: main
TEXSAMPLER: Attaching annotation to dict for node 0000022C97D5DF50
TEXSAMPLER: Adding DescriptorSet=0 annotation
TEXSAMPLER: Attaching annotation to dict for node 0000022C97D5DF50
TEXSAMPLER: Adding DescriptorBinding=1 annotation
TEXSAMPLER: Attaching annotation to dict for node 0000022C97D5DF50

IR dumped
TEXSAMPLER: Emitting SPIR-V for global at 0000022C97D5F7D0
TEXSAMPLER: Has 3 annotations
TEXSAMPLER: Emitting DescriptorSet decoration = 0
TEXSAMPLER: Emitting DescriptorBinding decoration = 1
Wrote result to output.spv

%105 = ref(UniformConstant, %104): CrossDevice %105 = @Exported("texSampler") @DescriptorSet(0) @DescriptorBinding(1) GlobalVariable(type: %104, address_space: UniformConstant, is_ref: true, init: null)
%116: CrossDevice %104 = Load(mem: %102, ptr: texSampler)

this is the llvm IR:

@.str.11 = private unnamed_addr constant [28 x i8] c"shady::builtin::LaunchIdKHR\00", section "llvm.metadata"
@.str.12 = private unnamed_addr constant [25 x i8] c"shady::descriptor_set::0\00", section "llvm.metadata"
@.str.13 = private unnamed_addr constant [28 x i8] c"..\\test\\vcc\\textured.frag.c\00", section "llvm.metadata"
@.str.14 = private unnamed_addr constant [29 x i8] c"shady::descriptor_binding::1\00", section "llvm.metadata"
@.str.15 = private unnamed_addr constant [15 x i8] c"shady::io::398\00", section "llvm.metadata"
@.str.16 = private unnamed_addr constant [19 x i8] c"shady::location::0\00", section "llvm.metadata"
@.str.17 = private unnamed_addr constant [19 x i8] c"shady::location::1\00", section "llvm.metadata"

this is the spirv disassembly:

               OpName %texSampler "texSampler"
               OpName %texSampler "texSampler"
               OpName %texSampler "texSampler"
               OpName %stack_ptr_2 "stack_ptr"
               OpName %generated_fini "generated_fini"
               OpName %fragColor_physical_2 "fragColor_physical"
               OpName %fragTexCoord_physical_2 "fragTexCoord_physical"
               OpName %outColor_physical_2 "outColor_physical"
               OpName %stack_ptr_3 "stack_ptr"
               OpName %generated_fini_0 "generated_fini"
               OpName %outColor "outColor"
               OpName %outColor "outColor"
               OpName %outColor "outColor"
               OpName %stack_ptr_4 "stack_ptr"
               OpDecorate %fragTexCoord Location 1
               OpDecorate %fragColor Location 0
               OpDecorate %texSampler DescriptorSet 0
               OpDecorate %texSampler Binding 1

that also has unrelated out and in variables

alright, next thing is learning how to create my own annotation and IR node, and have it emit spirv, just to learn

this is interesting though from the spirv dis

%41 = OpLoad %40 %texSampler
%42 = OpLoad %v2float %fragTexCoord_physical_1
%43 = OpImageSampleImplicitLod %v4float %41 %42

the fuck is code motion

[astral hinge]

the compiler can legally move code around if it doesn't affect its semantics

[cloud rivet]

ahh right

[astral hinge]

that's just clarifying that the compiler won't move your code in certain cases

[cloud rivet]

thanks makes sense

[astral hinge]

to be clear, it's talking about texture instructions that use implicit derivatives, which are undefined in non-uniform control flow